Claude Science: What It Is, How It Works, and Who It's For

Claude Science is a beta desktop application from Anthropic built for researchers rather than developers. It runs analyses, searches scientific databases, renders proteins and genomic tracks natively, and keeps a traceable record of the code and conversation behind every figure it produces. It downloads for Mac (Apple Silicon and Intel) and Linux from the Claude Science product page.

If you have used Claude Code, the shape is familiar: an agent with a persistent workspace and real execution. The difference is what the workspace holds. Claude Code's unit of work is a repository. Claude Science's unit of work is an analysis, and the thing it hands back is a figure, a table, a notebook, or a manuscript section.

What Claude Science actually is

It is a local application that manages the compute, data, and provenance around scientific analysis. Three things distinguish it from a general chat assistant that can run Python.

First, it renders scientific objects directly. Proteins, sequence alignments, genomic tracks, chemical structures, and PDFs display inside the app instead of being described in prose or dumped as coordinates.

Second, every artifact carries its history. A figure ships with the exact code, the environment it ran in, and the conversation that produced it. That makes a result auditable months later, when you no longer remember which normalisation you used.

Third, it holds state. Python and R kernels persist across a session, so variables survive between questions the way they do in a notebook you have been living in all afternoon.

How it works

Compute it manages for you

The app sets up and manages environments across three tiers: your laptop, a Linux machine, and an HPC login node. For cluster work it writes batch scripts and submits jobs over SSH, and Anthropic describes it scaling from a single GPU to hundreds.

This is the part that most "AI for science" tools skip. Getting a model to write a Slurm script is easy. Having it submit the job, watch it, and bring the output back into the same conversation is the tedious part, and it is the part that decides whether a tool survives contact with a real lab.

Kernels that remember

Persistent Python and R kernels keep variables alive across an analysis session. You load a dataset once, then ask five questions against it without re-reading from disk each time. R support matters more than it looks: a large share of bioinformatics and biostatistics tooling is R-first, and most coding agents treat R as an afterthought.

A reviewer running in the background

Claude Science runs a background reviewer over its own output. It flags incorrect citations, numbers it cannot trace back to a source, and figures that do not match the claim in the surrounding text.

Treat this as a first-pass filter, not a guarantee. It catches the class of error where an agent confidently cites a paper that says something adjacent to what it claimed. It does not catch a defensible analysis built on a bad assumption, and no reviewer that lives inside the same system that produced the work can be your only check.

Figures you edit by talking

Rather than regenerating a plot from scratch each time you want the axis relabelled, you annotate it in plain language and the app iterates on the figure. This is a small feature with a large effect on how many rounds of revision you are willing to do before a lab meeting.

Manuscript drafting in the same window

Analyses and prose live together, rendered as Markdown and LaTeX. The pitch is that the methods section is written next to the code that generated the numbers, so the two do not drift apart.

The databases and connectors

Claude Science comes pre-wired to more than 60 scientific databases, with literature integration included. Anthropic's Claude for Life Sciences page names several of the connectors in the wider platform, including 10x Genomics Cloud, Benchling, PubMed, bioRxiv, and ClinicalTrials.gov.

The practical read: the connector list is doing the work that a bespoke integration project would otherwise do. If your data lives in Benchling and your literature search lives in PubMed, a large fraction of the plumbing is already built. If your data lives in a LIMS nobody outside your institute has heard of, you are writing a custom connector, which the app supports.

What you can do with it today

Anthropic lists four workflows the app is set up for:

• Single-cell RNA-seq analysis — the clustering, annotation, and differential-expression loop.
• Phylogenetic and evolutionary analysis — alignment, tree building, and the interpretation around it.
• Protein structure and protein language model work — structural inspection alongside model-based prediction.
• Cheminformatics — molecule handling, property work, and chemical structure rendering.

Beyond those, pipelines can be saved as reusable skills, so a normalisation-and-QC sequence you have settled on becomes something you invoke rather than re-explain. If you have followed how Claude records a skill in the wider Claude ecosystem, this is the same idea aimed at analysis pipelines instead of office workflows.

Claude Science vs Claude for Life Sciences

These are two different things with confusingly similar names, and the distinction is worth getting right before you pitch either one internally.

Claude for Life Sciences is the broader platform and solution area. It runs in the browser through Claude, connects to cloud scientific databases, and is positioned around citations and audit trails for research teams.

Claude Science is the desktop application. It is where local compute, persistent kernels, native scientific renderers, and HPC job submission live.

If someone on your team says "we already have Claude for Life Sciences," they may mean browser access to connectors, which is not the same as the local execution environment.

Who it is for

It fits computational biologists, bioinformaticians, structural biologists, and cheminformaticians who already work in Python or R and are currently gluing together a notebook, a cluster, a literature manager, and a figure tool by hand.

It fits less well in three cases. Wet-lab-only groups with no computational component get little from an execution environment. Teams in regulated or IP-sensitive settings need to answer where analysis data goes before installing a beta, and the product page does not settle that question. And anyone who needs a stable, versioned, citable pipeline for a submission should be cautious about building on beta software whose pricing has not been published.

What is still unclear

Three honest gaps at the time of writing. Pricing and plan requirements are not stated on the product page. The relationship between the desktop app's local execution and the cloud connectors is not spelled out in detail, which matters for anyone with data-residency obligations. And "beta" is doing real work in that sentence: feature behaviour and availability can change.

Anthropic set the context for the launch in its Expanding our support for scientists announcement in late August 2026, alongside a broader push into research tooling.

Where this sits next to an open stack

Claude Science is a strong, opinionated bundle: renderers, connectors, compute management, and provenance in one closed application. The trade is the usual one. You get an integrated environment quickly, and you accept a single vendor for the model, the orchestration, and the data path.

If the constraint you are working under is that analysis data cannot leave your infrastructure, or that you need to run against local models, that trade may not be available to you. We compared the options in the best open-source Claude Science alternatives, and if you are trying to work out whether your team actually needs the science app or the coding one, Claude Science vs Claude Code walks through the split.

Frequently asked questions

Is it free? Anthropic has not published pricing or plan requirements on the product page. Assume it is tied to a paid Claude plan until stated otherwise, and confirm before budgeting.

Does it run on Windows? No. The downloads listed are Mac on Apple Silicon, Mac on Intel, and Linux.

Is it the same as Claude for Life Sciences? No. Life Sciences is the browser-based platform and solution area; the desktop app is where local execution, persistent kernels, and native renderers live.

Does it support R? Yes. It runs persistent Python and R kernels that keep variables alive across an analysis session, which is unusual among coding agents.

Can it use my cluster? Yes. It manages environments on HPC login nodes, writes batch scripts, and submits jobs over SSH.

Where does my data go? The app runs locally but the connectors reach cloud databases. If you have data-residency obligations, get that boundary in writing before you install it.

Run the same workflows on an open stack

Eigent is an open-source desktop AI workforce that runs multi-agent workflows locally, connects to MCP tools, and works with the model you choose, including local inference through Ollama or vLLM. For research teams, the useful pattern is the same one the app formalises: turn a pipeline you repeat into something the agents run for you, which is exactly what saving workflows as skills is for. Download Eigent and try it against a dataset you can already share.